GCN is basically found on cards starting with the 77xx and up. Does anyone know if those cards give more consistent performance than the older VLIW ones?

At the risk of giving bad advice, my HD 7770 worked fairly consistently that I recall. But my GTX 650 Ti was a little better (maybe 10 to 20 percent). However, I no longer have either card installed, so can't check it now. (Win7 64-bit for both.)

But I presently use an HD 7870 on Folding, and the new drivers (13.8 Beta) are a small improvement over the last ones (13.6 Beta2), which in turn are better than 13.4. In general, you want the latest AMD drivers due to the improvements in OpenCL.

I'm not sure I've ever seen such a calm NV <-> ATI thread, anywhere, ever. It almost restores my faith in humanity. Congratulations to everyone involved.

Thanks, i realy apreciate your msg.

---

For those of you who are having an easy time with ATI, here are some recent results of mine:


Run Time CPU Time Credit
1500 360 110
1600 310 110
1700 320 110
4300 80 100
5200 330 110
9100 290 100
9200 270 100
10900 280 110
10700 290 100
10700 280 100
4300 50 40
11900 310 110
4800 40 40

---

The latest AMD drivers (13.8 Beta) specifically mention "Improved stability for AMD Radeon HD 7790". They might help.

---

The latest AMD drivers (13.8 Beta) specifically mention "Improved stability for AMD Radeon HD 7790". They might help.

That is likely referring to graphics performance and not compute. AMD's driver release notes rarely say much about the OpenCL component, what's changed or what's been fixed.

It's always worth trying the new drivers out just to see, if only because they *do* say to little about what to expect OpenCL-wise.

But you have to be ready to roll back the driver if you don't like what you see, not a trivial task. And it's helpful if you can run the test benches (MB and AP) to test the driver offline without endangering your production BOINC wu's.

But you have to be ready to roll back the driver if you don't like what you see, not a trivial task.

It is best to do a clean uninstall first. I always uninstall the drivers first from the Control Panel/Programs and Features (in Win7). Do not try to uninstall from the Device Manager. Then, after a reboot, I run Driver Fusion to get rid of the remnants. You can use Driver Sweeper instead, but Driver Fusion usually finds a few more entries. Also, if you have ever had an Nvidia card installed before, it is best to get rid of its remnants too.

Finally, I install the new version.

But you have to be ready to roll back the driver if you don't like what you see, not a trivial task.

It is best to do a clean uninstall first. I always uninstall the drivers first from the Control Panel/Programs and Features (in Win7). Do not try to uninstall from the Device Manager. Then, after a reboot, I run Driver Fusion to get rid of the remnants. You can use Driver Sweeper instead, but Driver Fusion usually finds a few more entries. Also, if you have ever had an Nvidia card installed before, it is best to get rid of its remnants too.

Finally, I install the new version.

AFAIK, Driver Fusion/Sweeper does NOT remove OpenCL. The only confirmed products that Remove AMD OpenCL are 'AMD Catalyst Un-install Utility', which only works in Windows 7, and 'ATIman Uninstaller'. If you have Windows 8, as myself, your only choice is ATIman Uninstaller.

If you try to downgrade the ATI driver without using one of those two products, you will most likely have driver problems...

AFAIK, Driver Fusion/Sweeper does NOT remove OpenCL. The only confirmed products that Remove AMD OpenCL are 'AMD Catalyst Un-install Utility', which only works in Windows 7, and 'ATIman Uninstaller'. If you have Windows 8, as myself, your only choice is ATIman Uninstaller.

If you try to downgrade the ATI driver without using one of those two products, you will most likely have driver problems...

I never have problems on Win7. It is the combination procedure that I cited that gets rid of OpenCL. The problem with the AMD uninstallers is that they only remove AMD components. But a lot of the problems are due to leftover Nvidia entries, which Driver Fusion and Driver Sweeper can get rid of.

---

If you search the Net, you will find plenty of people complaining that the Free version of Driver Sweeper will NOT remove OpenCL. I've heard the PAID version will. I refuse to pay to have a Free version removed... Driver Fusion did Not work for Me using Windows 8. ATIman Uninstaller worked.

Here's a recent thread on a new Uninstaller, Display Driver Uninstaller Thread

-Avoid known conflict like the recent one with AMD and their OpenCL driver resulting in a crash with some OpenCL application. (This part has not been implemented yet)

(This part has not been implemented yet), typical...

The latest AMD drivers (13.8 Beta) specifically mention "Improved stability for AMD Radeon HD 7790". They might help.

That is likely referring to graphics performance and not compute. AMD's driver release notes rarely say much about the OpenCL component, what's changed or what's been fixed.

It's always worth trying the new drivers out just to see, if only because they *do* say to little about what to expect OpenCL-wise.

But you have to be ready to roll back the driver if you don't like what you see, not a trivial task. And it's helpful if you can run the test benches (MB and AP) to test the driver offline without endangering your production BOINC wu's.

Dont forget you have to delete the cached binaries in your projects folder after driver update.
Otherwise old binaries will be used and new drivers dont take much effect.

Those are
MultiBeam_Kernels_r1874.clHD5_GPU_Name.bin_V7
r1874_CPU_Name.wisdom and all files ending .bin.

Evenso there is something not everybody noticed until now.
For better performance you need to add -sbs switch in comandline either appinfo.xml or mb_cmdline_win_x86_SSE_OpenCL_ATi.txt.

Its set to 64 by default because of compatibility.
Cards with 1 GB RAM should use -sbs 256.
Speedup is up to 15%.

---

With each crime and every kindness we birth our future.

So, I installed the pre-HD4XXX version from Lunatics, and that seemed to speed up the slow jobs by 10-15%. However, there were still fast and slow jobs. I then upgraded the drivers to 13.8 beta, and deleted the cached binaries and wisdom files. BOINC or Seti@home then removed all my GPU multibeam tasks and started downloading a whole mess of astropulse.

When it starts, I get 8 copies of the message "App version needs CAL but GPU doesn't support it". I am guessing that this message is coming from the CPU multibeam tasks, as there are 8 cores.

So, is this working better now? Hard to say.

---

TJL didn't you say you have a 7790? Why are you using the HD4xxx Lunatics version? To try to smooth out the run times?

I'm not sure how well the latest drivers work with the old CAL apps. I know 13.8 gives you CAL version 1.4.1848, my startup logs show it, but maybe it doesn't play well with the older not-maintained apps.

Try the Lunatics readme files, they may shed light.

TJL didn't you say you have a 7790? Why are you using the HD4xxx Lunatics version? To try to smooth out the run times?

I'm not sure how well the latest drivers work with the old CAL apps. I know 13.8 gives you CAL version 1.4.1848, my startup logs show it, but maybe it doesn't play well with the older not-maintained apps.

Try the Lunatics readme files, they may shed light.

Actually, he said, "pre-HD4XXX version from Lunatics". That probably means he tried the Old Brook+ App on a Brand New 7790 that has only one driver release in it's history. When I look for a driver for the 7790, I'm sent to a 'Special' version, AMD Radeonâ„¢ HD7790 Catalystâ„¢ Performance Driver. Considering it's a New Card, I'm not surprised there may be issues with the first few round of drivers. The same thing happened with the 7750. It was a few releases before the bugs went away. The 7750 I have even had a couple Firmware Updates within the first year. The last update seemed to speed up AP tasks by around 5-10%.

Considering it's a new card, I'd use the Uninstall Utility to revert back to 13.4 and wait for the next regular Catalyst update. It usually takes a few releases before new cards work 'normally'. I'd also refrain from using Old Apps that weren't designed for the Brand New Cards. The Newest MB7 App is here, Mike's Page

The 7790-specific driver is from April, the fixes in it must have been folded into 13.4. The release notes for 13.4 include applicable cards:

AMD Radeon HD 7700 Series (incl. HD7790)

and later on it mentions the 7790-specific fixes for TressFX.

Whether or not to try 13.8 beta comes with all the caveats of beta software.

But if TJL has already installed 13.8 beta it might be simpler to keep that driver and install the latest Lunatics apps. I'm running 13.8 with no issues. If it doesn't work for him then he can do the rollback to 13.4.

Some points:

Yes, I was using the HD 4XXX series app to try and smooth things out. The pre-HD 4XXX line was a mistake. (Specifically, I am using the Win64 Lunatics 0.541 MB7_win)x86-SSE_OpenCL_ATi_t1843 app)

It seems to be marginally faster than HD 5XXX, but does not really smooth things. I still have fast jobs and slow jobs. "Fast" jobs take the same length as on a GTS450 and "slow" jobs are up to a factor ten slower.

The sequence of events was

Uninstall the 12.x drivers I was using
Reboot to VGA
Install 13.8
Reboot
Remove Wisdom files
Reinstall Lunatics apps

I have not freed a CPU core. However, the GPU s@h task is using <1% of the CPU according to the task manager. On other hosts, where I have freed a core, once could see that the GPU app was consuming a lot of CPU.

The GPU usage is very erratic. The GPU core frequency moves from 500 to 700 GHz, and the GPU activity moves from 0 to 96%, according to Catalyst. The time scale is seconds.

---

I have not freed a CPU core. However, the GPU s@h task is using <1% of the CPU according to the task manager. On other hosts, where I have freed a core, once could see that the GPU app was consuming a lot of CPU.

There's your problem right there. It is very counter-intuitive that even though the GPU job uses a few percent of a CPU core, it *still* needs the whole core all to itself. But it does, that is a well established empirical fact.

The theory is that there is some GPU<->host communication issue that is very timing-sensitive, to the nanosecond level. So the core has to be ready to respond to the GPU instantaneously.

Your GPU should be running at > 85% continuously. Free a core, I almost guarantee that is your answer.

I have not freed a CPU core. However, the GPU s@h task is using <1% of the CPU according to the task manager. On other hosts, where I have freed a core, once could see that the GPU app was consuming a lot of CPU.

There's your problem right there. It is very counter-intuitive that even though the GPU job uses a few percent of a CPU core, it *still* needs the whole core all to itself. But it does, that is a well established empirical fact.

The theory is that there is some GPU<->host communication issue that is very timing-sensitive, to the nanosecond level. So the core has to be ready to respond to the GPU instantaneously.

Your GPU should be running at > 85% continuously. Free a core, I almost guarantee that is your answer.

Exactly the problem with AMD/ATi cards, you must free core/s to run these cards, but it would help a lot more if you would un-hide your computers (no one can see sensitive info and if you don't believe me then look at mine, it will take an hour or 2 after un-hiding it/them for us to see your task details) as this would make things easier for us to help you.

If you run multiple tasks on an AMD/ATi card then you will need to free another core for each extra task to get the best performance.

Cheers.

Exactly the problem with AMD/ATi cards, you must free core/s to run these cards...

Not having ever run with NV I don't know by experience, but from what I read, people use idle cores with NV cards too. Is that not always the case?

Maybe it's only when running OpenCL?

Exactly the problem with AMD/ATi cards, you must free core/s to run these cards...

Not having ever run with NV I don't know by experience, but from what I read, people use idle cores with NV cards too. Is that not always the case?

Maybe it's only when running OpenCL?

In fact it depends on either the CPU and GPU.

I see a lot of NV hosts suffering on that as well.

---

With each crime and every kindness we birth our future.

The theory is that there is some GPU<->host communication issue that is very timing-sensitive, to the nanosecond level. So the core has to be ready to respond to the GPU instantaneously

That theory would suggest that once one core is free, one can run as many GPU instances as one would like, at least until the freed core fills up.

So, I did two things: I freed a core (an HT-pseudocore, actually) and yes, the usage jumped to 95%. On Astropulse, which is what's running now, it's 8x faster than it was. Thanks!

I reverted to 13.4. Note that for a 7790, 12.101 is the recommended version. That stopped the CAL messages, and restored the Multibeam tasks in job list. They are labeled opencl_ati_sah, whereas Astropulse is ati_opencl_100.

Getting back to the original question, I would argue that ATI is certainly less out-of-the-box than Nvidia. It requires editing app_info.xml to use it effectively. The relative performance differences are mitigated by the requirement of freeing a core. This card is getting 700 credits/hour on Astropulse, vs. 550/hour for a GTS450. However, it requires freeing up a core, and one CPU core on Astropulse generates 105/hour. So the gain is really only 45 credits/hour - an 8% boost when the Boinc benchmarking would suggest a factor of 4. (2400 GFlops vs. 600)

---